Omkar1806 commited on
Commit
7eb8ba2
·
verified ·
1 Parent(s): 603d2de

Update env.py

Browse files
Files changed (1) hide show
  1. env.py +54 -86
env.py CHANGED
@@ -2,9 +2,12 @@ import numpy as np
2
  import gymnasium as gym
3
  from gymnasium import spaces
4
 
5
- # ---------------------------------------------------------------------------
6
- # Data & Vocabulary
7
- # ---------------------------------------------------------------------------
 
 
 
8
  KEYWORD_VOCAB = [
9
  "invoice", "payment", "overdue", "refund", # billing
10
  "hacked", "breach", "unauthorized", "password", # security
@@ -18,111 +21,76 @@ SENTIMENT_MAP = {"positive": 0, "neutral": 1, "negative": 2}
18
  CONTEXT_MAP = {"spam": 0, "billing": 1, "tech": 2, "security": 3, "legal": 4}
19
  OBS_DIM = len(KEYWORD_VOCAB) + len(SENTIMENT_MAP) + len(CONTEXT_MAP)
20
 
21
- EMAIL_DATASET = [
22
- {"id": "TC-H-01", "description": "IT password reset disguised", "keywords": ["password", "unauthorized", "critical", "urgent"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
23
- {"id": "TC-H-02", "description": "Fake invoice portal redirect", "keywords": ["invoice", "payment", "password", "unauthorized"], "sentiment": "neutral", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
24
- {"id": "TC-H-03", "description": "HR benefits credential capture", "keywords": ["password", "urgent", "critical"], "sentiment": "neutral", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
25
- {"id": "TC-H-04", "description": "Account suspension notice", "keywords": ["unauthorized", "password", "breach", "urgent"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
26
- {"id": "TC-H-09", "description": "Fake GDPR breach notice", "keywords": ["breach", "password", "legal", "threat"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
27
- {"id": "TC-H-10", "description": "Ransomware / Compliance audit", "keywords": ["hacked", "breach", "lawsuit", "legal"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 2, 2)},
28
- {"id": "TC-H-11", "description": "Extortion threat", "keywords": ["hacked", "breach", "attorney", "threat"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 2, 2)},
29
- {"id": "TC-H-12", "description": "Law firm data misuse letter", "keywords": ["unauthorized", "breach", "attorney", "lawsuit"], "sentiment": "negative", "context": "legal", "difficulty": "hard", "correct_actions": (2, 2, 2)},
30
- {"description": "Account hacked - urgent", "keywords": ["hacked", "unauthorized", "password"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
31
- {"description": "Data breach notification", "keywords": ["breach", "unauthorized", "critical"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 1, 2)},
32
- {"description": "Legal threat - lawsuit", "keywords": ["lawsuit", "legal", "attorney", "threat"], "sentiment": "negative", "context": "legal", "difficulty": "hard", "correct_actions": (2, 2, 2)},
33
- {"description": "Ransomware threat", "keywords": ["hacked", "threat", "critical", "breach"], "sentiment": "negative", "context": "security", "difficulty": "hard", "correct_actions": (2, 2, 2)},
34
- ]
35
-
36
- # ---------------------------------------------------------------------------
37
- # Environment
38
- # ---------------------------------------------------------------------------
39
  class EmailTriageEnv(gym.Env):
40
- def __init__(self, task="hard"):
41
  super().__init__()
42
- self.email_batch = [e for e in EMAIL_DATASET if e.get("difficulty") == task]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43
  self.action_space = spaces.MultiDiscrete([3, 3, 3])
44
  self.observation_space = spaces.Box(low=0.0, high=1.0, shape=(OBS_DIM,), dtype=np.float32)
45
  self._step_idx = 0
46
 
47
  def _encode(self, email):
48
- kw_flags = np.array([1.0 if kw in email["keywords"] else 0.0 for kw in KEYWORD_VOCAB])
49
- sentiment_vec = np.zeros(len(SENTIMENT_MAP)); sentiment_vec[SENTIMENT_MAP[email["sentiment"]]] = 1.0
50
- context_vec = np.zeros(len(CONTEXT_MAP)); context_vec[CONTEXT_MAP[email["context"]]] = 1.0
51
- return np.concatenate([kw_flags, sentiment_vec, context_vec])
 
 
 
 
52
 
53
  def reset(self, seed=None, options=None):
54
  super().reset(seed=seed)
55
  self._step_idx = 0
56
- return self._encode(self.email_batch[0]), {"description": self.email_batch[0]["description"]}
 
 
 
 
57
 
58
  def step(self, action):
59
  email = self.email_batch[self._step_idx]
60
  correct = email["correct_actions"]
 
 
61
  reward = 0.0
62
-
63
- # Critical: Penalty for missing Security Urgency
64
  if correct[0] == 2 and action[0] != 2:
65
- reward = -2.0
66
  elif tuple(action) == correct:
67
  reward = 1.0
68
  elif action[0] == correct[0]:
69
  reward = 0.2
70
 
71
  self._step_idx += 1
72
- done = self._step_idx >= len(self.email_batch)
73
- obs = self._encode(email) if not done else np.zeros(OBS_DIM)
74
 
75
- return obs, float(reward), done, False, {"description": email["description"], "correct": correct}
76
-
77
- # ---------------------------------------------------------------------------
78
- # Smart Agent
79
- # ---------------------------------------------------------------------------
80
- class EmailGatekeeperAgent:
81
- def predict(self, obs):
82
- is_security_kw = any(obs[4:8]) # hacked, breach, unauthorized, password
83
- is_legal_kw = any(obs[12:16]) # lawsuit, legal, attorney, sue
84
- context_security = obs[28] == 1.0
85
- context_legal = obs[29] == 1.0
86
-
87
- # Default: General
88
- urgency, routing, resolution = 0, 0, 0
89
-
90
- # Rule 1: High Urgency for Security/Legal
91
- if context_security or context_legal or is_security_kw:
92
- urgency = 2 # Security Breach Level
93
- resolution = 2 # Escalate to Human
94
-
95
- # Rule 2: Routing Logic
96
- # Agar Legal context hai ya specific keywords hain, toh Legal routing
97
- if context_legal or is_legal_kw:
98
- routing = 2 # Legal
99
- else:
100
- routing = 1 # Tech Support (Security/Phishing cases)
101
 
102
- return np.array([urgency, routing, resolution])
103
-
104
- # ---------------------------------------------------------------------------
105
- # Main Execution
106
- # ---------------------------------------------------------------------------
107
- if __name__ == "__main__":
108
- env = EmailTriageEnv(task="hard")
109
- agent = EmailGatekeeperAgent()
110
- obs, info = env.reset()
111
-
112
- total_reward = 0
113
- print(f"{'Email Description':<40} | {'Agent Action':<12} | {'Reward'}")
114
- print("-" * 70)
115
-
116
- done = False
117
- count = 0
118
- while not done:
119
- action = agent.predict(obs)
120
- obs, reward, done, _, info = env.step(action)
121
- total_reward += reward
122
- count += 1
123
- print(f"{info['description'][:38]:<40} | {str(action):<12} | {reward:>6}")
124
-
125
- final_score = max(0, total_reward / count)
126
- print("-" * 70)
127
- print(f"TOTAL SCORE: {final_score:.3f}")
128
-
 
2
  import gymnasium as gym
3
  from gymnasium import spaces
4
 
5
+ # --- UI LABELS (Ye labels app.py mang raha hai) ---
6
+ URGENCY_LABELS = ["General", "Billing", "Security Breach"]
7
+ ROUTING_LABELS = ["AI Auto-Reply", "Tech Support", "Legal"]
8
+ RESOLUTION_LABELS = ["Archive", "Draft Reply", "Escalate to Human"]
9
+
10
+ # --- Vocabulary & Encoding Configuration ---
11
  KEYWORD_VOCAB = [
12
  "invoice", "payment", "overdue", "refund", # billing
13
  "hacked", "breach", "unauthorized", "password", # security
 
21
  CONTEXT_MAP = {"spam": 0, "billing": 1, "tech": 2, "security": 3, "legal": 4}
22
  OBS_DIM = len(KEYWORD_VOCAB) + len(SENTIMENT_MAP) + len(CONTEXT_MAP)
23
 
24
+ # --- Environment Class ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25
  class EmailTriageEnv(gym.Env):
26
+ def __init__(self, task="all", batch=None, shuffle=True):
27
  super().__init__()
28
+
29
+ # Dataset ko import karna (app.py se load hoga)
30
+ try:
31
+ from app import EMAIL_DATASET
32
+ dataset_to_use = EMAIL_DATASET
33
+ except ImportError:
34
+ dataset_to_use = [] # Fallback agar dataset na mile
35
+
36
+ if batch is not None:
37
+ self.email_batch = batch
38
+ elif task != "all":
39
+ self.email_batch = [e for e in dataset_to_use if e.get("difficulty") == task]
40
+ else:
41
+ self.email_batch = dataset_to_use
42
+
43
+ self.shuffle = shuffle
44
  self.action_space = spaces.MultiDiscrete([3, 3, 3])
45
  self.observation_space = spaces.Box(low=0.0, high=1.0, shape=(OBS_DIM,), dtype=np.float32)
46
  self._step_idx = 0
47
 
48
  def _encode(self, email):
49
+ kw_flags = np.array([1.0 if kw in email.get("keywords", []) else 0.0 for kw in KEYWORD_VOCAB])
50
+ sent_idx = SENTIMENT_MAP.get(email.get("sentiment", "neutral"), 1)
51
+ sentiment_vec = np.zeros(len(SENTIMENT_MAP)); sentiment_vec[sent_idx] = 1.0
52
+
53
+ ctx_idx = CONTEXT_MAP.get(email.get("context", "spam"), 0)
54
+ context_vec = np.zeros(len(CONTEXT_MAP)); context_vec[ctx_idx] = 1.0
55
+
56
+ return np.concatenate([kw_flags, sentiment_vec, context_vec]).astype(np.float32)
57
 
58
  def reset(self, seed=None, options=None):
59
  super().reset(seed=seed)
60
  self._step_idx = 0
61
+ if not self.email_batch:
62
+ return np.zeros(OBS_DIM, dtype=np.float32), {}
63
+
64
+ obs = self._encode(self.email_batch[0])
65
+ return obs, {"description": self.email_batch[0].get("description", "")}
66
 
67
  def step(self, action):
68
  email = self.email_batch[self._step_idx]
69
  correct = email["correct_actions"]
70
+
71
+ # Reward Logic (Score sudharne ke liye)
72
  reward = 0.0
 
 
73
  if correct[0] == 2 and action[0] != 2:
74
+ reward = -2.0 # Security missed penalty
75
  elif tuple(action) == correct:
76
  reward = 1.0
77
  elif action[0] == correct[0]:
78
  reward = 0.2
79
 
80
  self._step_idx += 1
81
+ terminated = self._step_idx >= len(self.email_batch)
 
82
 
83
+ # Next observation
84
+ if not terminated:
85
+ next_email = self.email_batch[self._step_idx]
86
+ obs = self._encode(next_email)
87
+ else:
88
+ obs = self._encode(email)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
 
90
+ info = {
91
+ "description": email.get("description", ""),
92
+ "correct_actions": correct,
93
+ "raw_reward": reward
94
+ }
95
+
96
+ return obs, float(reward), terminated, False, info